Robert found a case when the ordinal map can be quite wasteful in terms of memory usage: in order to be able to resolve values given a global ordinals, it stores two things:
- an identifier of the segment where the value is
- the difference between the ordinal on the segment and the global ordinal
The issue is that OrdinalMap currently picks any of the segments that contain the value but we can do better: we can pick the first segment that has the value. This will help for two reasons:
- it will potentially require fewer bits per value to store the segment ids if NRT segments don't introduce new values
- if all values happen to appear in the first segment, then the map from global ords to deltas only stores zeros.
I just tested on an index where all values are in the first segment and this helped reduce memory usage of the ordinal map by 4x (from 3.5MB to 800KB).
|Field||Original Value||New Value|
|Fix Version/s||4.10 [ 12327121 ]|
|Fix Version/s||4.9 [ 12326730 ]|
|Status||Open [ 1 ]||Resolved [ 5 ]|
|Resolution||Fixed [ 1 ]|